Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Adaptive Split-MMD Training for Small-Sample Cross-Dataset P300 EEG Classification

Weiyu Chen, Arnaud Delorme

TL;DR

This work tackles cross-dataset transfer in single-trial P300 EEG under a strict small-sample regime, where target trials are scarce (e.g., $10$ per subject) and domain shift impedes learning. It introduces Adaptive Split–MMD Training (AS-MMD), a backbone-agnostic recipe that combines a target-weighted loss with a warm-up schedule, Split‑BN normalization, and a parameter-free logit-level RBF-MMD penalty, implemented on a compact EEG Conformer. Empirical results across two public datasets (AVO and ERP CORE P3) show AS-MMD consistently improves target-domain accuracy and AUC over target-only and pooling baselines in both transfer directions, with ablations confirming contributions from all components. The method maintains the same inference-time architecture, requires no extra trainable modules, and offers a practical default for deploying cross-dataset, data-limited EEG decoders in real-world settings.

Abstract

Detecting single-trial P300 from EEG is difficult when only a few labeled trials are available. When attempting to boost a small target set with a large source dataset through transfer learning, cross-dataset shift arises. To address this challenge, we study transfer between two public visual-oddball ERP datasets using five shared electrodes (Fz, Pz, P3, P4, Oz) under a strict small-sample regime (target: 10 trials/subject; source: 80 trials/subject). We introduce Adaptive Split Maximum Mean Discrepancy Training (AS-MMD), which combines (i) a target-weighted loss with warm-up tied to the square root of the source/target size ratio, (ii) Split Batch Normalization (Split-BN) with shared affine parameters and per-domain running statistics, and (iii) a parameter-free logit-level Radial Basis Function kernel Maximum Mean Discrepancy (RBF-MMD) term using the median-bandwidth heuristic. Implemented on an EEG Conformer, AS-MMD is backbone-agnostic and leaves the inference-time model unchanged. Across both transfer directions, it outperforms target-only and pooled training (Active Visual Oddball: accuracy/AUC 0.66/0.74; ERP CORE P3: 0.61/0.65), with gains over pooling significant under corrected paired t-tests. Ablations attribute improvements to all three components.

Adaptive Split-MMD Training for Small-Sample Cross-Dataset P300 EEG Classification

TL;DR

This work tackles cross-dataset transfer in single-trial P300 EEG under a strict small-sample regime, where target trials are scarce (e.g., per subject) and domain shift impedes learning. It introduces Adaptive Split–MMD Training (AS-MMD), a backbone-agnostic recipe that combines a target-weighted loss with a warm-up schedule, Split‑BN normalization, and a parameter-free logit-level RBF-MMD penalty, implemented on a compact EEG Conformer. Empirical results across two public datasets (AVO and ERP CORE P3) show AS-MMD consistently improves target-domain accuracy and AUC over target-only and pooling baselines in both transfer directions, with ablations confirming contributions from all components. The method maintains the same inference-time architecture, requires no extra trainable modules, and offers a practical default for deploying cross-dataset, data-limited EEG decoders in real-world settings.

Abstract

Detecting single-trial P300 from EEG is difficult when only a few labeled trials are available. When attempting to boost a small target set with a large source dataset through transfer learning, cross-dataset shift arises. To address this challenge, we study transfer between two public visual-oddball ERP datasets using five shared electrodes (Fz, Pz, P3, P4, Oz) under a strict small-sample regime (target: 10 trials/subject; source: 80 trials/subject). We introduce Adaptive Split Maximum Mean Discrepancy Training (AS-MMD), which combines (i) a target-weighted loss with warm-up tied to the square root of the source/target size ratio, (ii) Split Batch Normalization (Split-BN) with shared affine parameters and per-domain running statistics, and (iii) a parameter-free logit-level Radial Basis Function kernel Maximum Mean Discrepancy (RBF-MMD) term using the median-bandwidth heuristic. Implemented on an EEG Conformer, AS-MMD is backbone-agnostic and leaves the inference-time model unchanged. Across both transfer directions, it outperforms target-only and pooled training (Active Visual Oddball: accuracy/AUC 0.66/0.74; ERP CORE P3: 0.61/0.65), with gains over pooling significant under corrected paired t-tests. Ablations attribute improvements to all three components.
Paper Structure (26 sections, 6 equations, 3 figures, 3 tables, 1 algorithm)

This paper contains 26 sections, 6 equations, 3 figures, 3 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Datasets and Task
  4. Dataset
  5. Dataset 1: Active Visual Oddball subset of "Cognitive Electrophysiology in Socioeconomic Context in Adulthood"
  6. Dataset 2: ERP CORE P3
  7. Task Setup
  8. Objective, label space, and event mapping
  9. Common electrodes and preprocessing
  10. Evaluation protocol
  11. Method
  12. Backbone
  13. Baselines
  14. Target-only
  15. Pooled (Source + Target)
  16. ...and 11 more sections

Figures (3)

  • Figure 1: Overview of AS-MMD on the EEG Conformer backbone.
  • Figure 2: Comparison of baselines and AS-MMD on AVO and ERP CORE P3.
  • Figure 3: Ablation study: distributions of accuracy and AUC for AS-MMD and ablations on AVO and ERP CORE P3.